AI-Flow.eu

Docs, diagrams, and API entry points.

A compact hub for architecture maps, interactive flow views, API documentation, and the developer reference.

Architecture

Visual system maps

Developer Resources

Core operational documentation

Benchmarks

How well does retrieval actually work?

We test retrieval with a fixed catalog of 25 questions against a 74-document knowledge base: fictional company documents in English and German, plus 60 decoy documents that reuse the same vocabulary. The run uses real embeddings, real Atlas vector search, and the exact pipeline this app runs in production. No LLM grades anything. A question counts as answered when the passage containing the answer shows up in the results. Last run: August 30, 2026.

Question type Questions Answer in top 5 Answer in top 10 MRR
One clear fact4100%100%1.00
Fact buried mid-document3100%100%1.00
Exact identifier lookup2100%100%1.00
Paraphrased question367%67%0.70
German question4100%100%1.00
Answer spans 2 documents, single search475%88%1.00
Answer spans 3 documents, single search1100%100%1.00
Fact in a linked appendix2100%100%1.00
Multi-document chains, agent follow-up search55 of 5 resolved, every hop ranked first

MRR is 1.00 when the right passage is always the very first result.

Single facts are basically solved. Whether the fact is prominent, buried in the middle of a document, hidden behind an exact ID like NX-4471-KRANICH, or asked in German, the right passage comes back first.

Questions whose answer spans multiple documents behave differently, and the table shows it honestly. One search reliably finds the first document, but the second one only reaches the top 10 for four of five chains. This is why agents search more than once: with a follow-up query, all five chains resolve with the right passage ranked first, including one that crosses three documents.

Two catalog questions have no answer in the knowledge base at all. Both came back with only floor-level scores, so the retriever does not fake an answer when there is none.

A full search takes 333 ms on average (p95: 537 ms): about 170 ms embedding the query, 75 ms in the database, and 90 ms of post-processing. A query costs 14 embedding tokens on average (measured from the API's usage field), and a search returns about 2,300 tokens of context to the model. Re-running the entire benchmark, ingest included, costs around 21,000 embedding tokens, well under a cent, and the freshly inserted chunks were searchable in the vector index after 5 seconds.

The whole benchmark, corpus, questions and runner, lives in the repository under tools/retrievalBenchmark, so the numbers can be reproduced at any time.